Your First Agent — Full Walkthrough
This is the extended version of the quickstart, covering every configuration option.
Creating an agent
Section titled “Creating an agent”- If this is your first agent, the Console home asks What should your first agent do? Describe the work and choose Draft my agent, or pick one of three ready agents.
- From your second agent on (Pro and up), go to Agents and choose New agent. The page asks the same question (Draft with AI), lists ready agents on four shelves (Engineering, Go-to-market, Operations and Business), and ends with Start blank.
- Check the details. A draft shows what the agent will do and what it needs; Change something opens its details. A ready agent’s card opens them too, and Start blank opens them empty:
- Name: something descriptive, like “Support bot” or “Dev assistant”
- What should this agent do?: one line that becomes the description your team sees
- Who will talk to it?: Your team or Your customers. See Agent types →
- Instructions: what the agent is told to do (shown here for a draft; a ready agent comes with its own, and after Start blank you write them on the agent’s page)
- Choose Create and try it. The agent’s first chat opens in the same tab, and an Internal agent’s computer is already starting. A ready agent’s first prompt is waiting in the box; nothing is sent until you send it. If member governance holds your agent for approval, you stay on its page until an owner or admin approves it.
Use this agent on a ready agent does steps 3 and 4 in one click. If the agent needs an Integration you haven’t connected, such as GitHub, it asks you to connect it first.
Configuring the agent
Section titled “Configuring the agent”The agent page is one page. A rail down the left side lists its sections, you move between them freely, and everything you change collects in a single Save bar at the bottom — nothing is written until you save it. A Try it pane on the right runs the agent so you can feel a change before it goes anywhere.
Instructions
Section titled “Instructions”The instructions are the most important part. Define:
- Who the agent is — its role and identity
- What it should do — the tasks it handles
- How it should behave — tone, format, what to do when it’s unsure
See Writing Agent Instructions → for detailed guidance and examples.
Which model thinks for this agent. How hard it thinks is Auto (the default, which chooses per message), Fast (quicker, lower cost) or Deep (more thorough).
Capabilities
Section titled “Capabilities”Capabilities are what the agent can do on its own — turn on only what this agent needs:
- Web search — search the live web
- Location & maps — ground answers in Google Maps places data
- Browser — navigate and act on web pages, including vision-based clicks and typing
- Coding — repo-aware code editing on the agent’s computer (paid plans)
- Presentations, Diagrams, Image generation, Audio generation and Video generation — create that kind of output
- Memory — the agent remembers context across separate conversations
- Chat input switches — what people can attach or capture in the chat composer
An Internal agent always has its own computer (a sandbox with a terminal and files), with no switch for it; customer-facing agents don’t get a sandbox or browser. See Sandboxes → and Computer Use & Browser →.
Integrations
Section titled “Integrations”Select which integrations this agent can use, and which Connection it acts through. Only enable what’s needed — this helps keep the agent focused and prevents accidental writes.
Knowledge
Section titled “Knowledge”Attach the knowledge collection the agent should be able to search. One collection per agent.
Skills
Section titled “Skills”Attach built-in Skills for the work your Agent does, from code review and research to finance, sales and support. See Skill Library →.
Autonomy
Section titled “Autonomy”How much this agent may do before it checks with a person, plus any allow / ask / deny rules for specific actions.
The rest of the rail
Section titled “The rest of the rail”Voice (whether people can talk to it, and what it sounds like), Channels (where customers reach it), Access (which teammates can use or edit it), Computer (where its sandbox runs), Activity, History (every saved change, with the option to bring one back) and Danger zone (deactivate or delete).
Testing the agent
Section titled “Testing the agent”Create and try it opens the agent’s first chat for you. To come back to it, choose Chat at the top of the agent’s page (it opens in a new tab), or Open your chats at the bottom of the sidebar. The Try it pane on the agent’s page runs a quick test without leaving the settings.
Things to test:
- Does it follow the instructions?
- Does it use the knowledge correctly?
- Does it use its Integrations and Capabilities when needed?
- What happens when you give it an edge case or an off-topic request?
Deploying the agent
Section titled “Deploying the agent”Once you’re happy with the agent in testing, deploy it (Channels, the website widget and API keys are on Pro and up):
- Web Widget — embed a chat bubble on your website
- Telegram — connect a bot via BotFather
- Slack — add as a workspace bot
- WhatsApp — through your Twilio account
- REST API — call from your own application
- Loops — run the agent on a schedule or when an event arrives
Monitoring runs
Section titled “Monitoring runs”After you deploy, open the Activity section on the agent page to see its runs, success rate, average response time and credits, and its most recent runs. Open a run, or choose View all for the Runs page, to see:
- What the agent did at each step and what came back
- The output it produced
- Credit usage for that run
This is your primary tool for debugging and improving the agent over time.